Assessing the performance and interpretability of the CNN-LSTM-Attention model for daily streamflow forecasting in typical basins of the eastern Qinghai-Tibet Plateau
Chuchu Zhang, Yuyan Zhou, Lu Fan, Jianwei Liu, Jiayue Zhang, Zeying Yin, Mengyi Ji, Baoqi Li
Abstract
Hydrological forecasting is of great significance to regional water resources management and reservoir operation. Climate change has increased the complexity and difficulty of hydrological forecasting. In this study, a hybrid explainable streamflow forecasting model based on CNN-LSTM-Attention was established for five typical river source regions in the eastern Qinghai-Tibet Plateau (EQTP). The model effectively simulates typical basins in the EQTP, achieving an NSE range of 0.79 to 0.92 and an R 2 range of 0.81 to 0.93, which is better than LSTM. Incorporating base flow as an input significantly improves high-flow results in all basins, with mixed flow basins showing greater optimization than single flow basins. Higher base flow, increased daily minimum temperatures, lower relative humidity, and higher precipitation positively impact the model’s simulation and prediction capabilities.